AI systems built inside e-commerce operations.
I embed with your team for four to six weeks and leave a working system in production — not a roadmap. Amazon, DTC, and brand portfolios. Remote, fixed scope, fixed price.
Apply for a scoping sprint- 15 years Embedded in other people’s systems. Deployed into enterprise accounts — UNEDIC, Total, Roche, Bayer, Suez Lyonnaise des Eaux — then a decade running my own firm as the outsourced IT director for around fifty companies.
- 13 years Selling on Amazon FBA in a real account, US and EU.
- 7 marketplaces Brands launched across eighteen product categories — FR, DE, IT, ES, UK, US, CA.
- In production An eight-agent pipeline I built and run. Free audit, no signup, ten minutes.
Most e-commerce teams have already tried AI. A pilot that impressed everyone in a demo and then sat there. The models were never the problem — the integration was. Your data lives in four systems that don’t talk, your ops team can’t maintain what an agency built, and nobody owns the result once the invoice is paid.
Who you would be working with
David Daddi.
Today
I build and run a multi-agent system in production on a brand I operate myself — the approval gates, the audit logging and the deployment discipline in the next section. It runs this month, and I wrote it.
Fifteen years running other people’s systems
1998 to 2013, in France. I started on rotation: eleven enterprise deployments between 1998 and 2003 — UNEDIC, Total, Roche, Bayer, Suez Lyonnaise des Eaux, and others across insurance, pharmaceuticals, energy, public sector and retail — scoped and stabilised inside the client’s own operation. I worked inside those environments; I did not sell them.
Then for ten years I was the outsourced IT department for around fifty companies across south-west France, from five-person firms to two-thousand-employee industrial groups. By the end my team and I ran about a hundred servers and a thousand daily users across concrete, plastics moulding, sheet metal, solar installation, real estate, call centres, vineyards, hotels, and franchise networks of bakeries and salons. No two of them had the same systems, the same data, or the same idea of what “urgent” meant.
That is the part that matters here. I have never had the luxury of a clean environment. Fifty businesses with nothing in common is the same problem as forty acquired brands sitting on one portfolio — and I ran it as a fractional practice ten years before anyone called it that. I built it into an eleven-person company and sold it in 2013.
The domain
Thirteen years selling on Amazon FBA in a real account, US and EU. Between 2019 and 2023 I worked alongside brands launching products across seven marketplaces — France, Germany, Italy, Spain, the UK, the US and Canada — and across categories that had almost nothing in common: supplements, cosmetics, luxury food, air purifiers, mosquito control, industrial fasteners, gaming accessories, barbecues, baby, pet, refurbished phones.
Several of those are regulated. That is why the system I build refuses to write to a regulated listing without a human approving it, and why category detection fails closed rather than guessing. Those guardrails are not caution in the abstract. I have seen what happens without them.
I know what a Seller Central export looks like at 2am, and what it costs when a listing goes down.
Evidence
What I can actually show you.
No client logos here yet, and I’m not going to manufacture any. Here is exactly where things stand.
Keoxs AIO is live. An eight-agent pipeline running in production, and you can run a free audit against your own listing in about ten minutes without talking to me first.
The governance model below is a different thing. It’s design work — third internal version, still on paper. I’m starting the build now, and it goes on a brand I own before it goes anywhere near someone else’s account. That order is deliberate, and I’ll be doing it in public.
I’ll walk through any of it live on a call.
An eight-agent pipeline, 22 tools, in production today.
I built it and I run it. It’s the first Amazon listing optimizer built on Amazon’s own published research rather than on scraping. Not a screenshot: run the free Forensic Audit on one of your own ASINs and judge the output yourself.
Run the free Forensic Audit at keoxs.com →Three autonomy levels, every spend gated by a human, and a default of refusing to act when data is missing. This is the governance model I am designing for agents that act on an account rather than just observe it. Tap the image to read it full size.
Running in production today, on a brand I operate myself:
- No agent writes to a live listing without a human approving it.
- Regulated categories are refused by default rather than guessed at.
- Every model call is logged with its inputs and its outputs.
- Every deployment is traceable to a build and a revision.
- Every architectural decision is written down with an ID and a reason.
The engagement
Six weeks, and what happens in each one.
The model is forward-deployed: I work inside your systems rather than alongside them. Your Slack, your Seller Central, your repo, your data. That means deployed in your stack — not sitting in your office. I’ve done this from Miami for teams in three time zones.
Six weeks, one workflow, end to end.
-
Week 1 Immersion
I sit behind the people doing the work. I time them. I come out with a number, not a vision. Access hunting starts on day one.
-
Week 2 Ontology
We name the objects in your language: Product, Listing, Variant, Asset, Claim. This is where half the chaos explains itself.
-
Week 3–4 Build
One workflow, end to end, in your environment.
Working demo at the end of week 3. Not a slide.
-
Week 5 Production & adoption
The test is not “it works”. The test is “your catalog manager used it on Thursday without me”.
-
Week 6 Measurement & expansion map
The week-1 number, measured again. Then the next three projects, costed.
Boundaries
What this is not.
- I don’t promise rankings, sales, or revenue outcomes.
- I don’t read or access Amazon’s live systems. What I run is a simulation calibrated on Amazon’s published research, and I’m explicit about what’s measured, what’s inferred, and what’s not.
- I don’t write to your listings without a human approval step.
- I don’t run your PPC, produce your content, or manage your agency.
- I don’t migrate your whole catalog in the first engagement. One workflow, end to end, done properly.
- I don’t hand off to a junior. Every engagement is me.
Fit
Who this works for.
Good fit
- An established e-commerce brand or portfolio.
- One operational workflow that visibly costs you time.
- Someone who can grant system access in week 1.
- A single decision-maker.
Not a fit
- You’re looking for a developer by the day.
- You want a tool without changing a process.
- There’s no budget until next quarter.
- You expect a guaranteed commercial result.
- You need someone on site full time.
Engagement
How this starts.
Every engagement begins with a paid scoping sprint: 1–2 weeks, $3,500 – $7,500 depending on the size of the account. I map the workflows, rank three use cases by return, and build a throwaway prototype against your real data. It is credited against the build if you go ahead.
I quote the build at the end of scoping, never before. Quoting an environment you haven’t inspected is how a client ends up paying for a scope nobody agreed on. Builds run 4–6 weeks at a fixed scope and a fixed price, in the tens of thousands. Ongoing work is monthly, priced against a system that exists rather than one we imagined.
Payment is staged: part on signature, part at the week-3 demo, part on production. If we agree on any on-site visits, those are billed separately.
Apply
Apply for a scoping sprint.
Application sent.
I read every submission personally. You’ll hear back within two business days — either with a first set of questions about the workflow you described, or with a straight no and the reason.